Improvement in Land Cover Classification Using Multitemporal Sentinel-1 and Sentinel-2 Satellite Imagery

For improving the performance of multitemporal Sentinel-1 (S1) SAR and Sentinel-2 (S2) optical imagery for land cover classification, a framework based on Rotating Kernel Transformation (RKT) denoising algorithm and the stratified sampling method based on the Crop Data Layer (CDL) is proposed. Random Forest classifications based on different denoising algorithms and sampling methods are carried out to compare their accuracy and applicability for land cover classification. The results show that the RKT algorithm and the stratified sampling method can significantly improve the classification accuracy. The classification accuracy by S1 data alone without denoising (overall accuracy: 0.873, Kappa: 0.796) is significantly lower than that of S2 (overall accuracy: 0.979, Kappa: 0.970) resulting from effects of serious salt-and-pepper noise. After RKT filtering, the speckle noise of the S1 classification result is significantly reduced and the accuracy is significantly improved (overall accuracy: 0.944, Kappa: 0.912). RKT filter outperforms the Lee and Median filters in improving the classification accuracy of SAR imagery. Feature-level fusion of S1 and S2 achieves the highest classification accuracy (overall accuracy: 0.983, Kappa: 0.972) which is significantly higher than that of S1 and slightly higher than that of S2 data alone. It proves that the fusion of the optical and SAR data can weaken the speckle noises on classification maps and improve the classification accuracy. The stratified sampling method applied in this study significantly improves the classification accuracy of each experimental group, with the overall accuracy increasing by about 10% and the Kappa coefficient increasing by more than 15% on average.